Ordered or Orderless: A Revisit for Video Based Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video-based person re-identification (VPRe-id) commonly uses recurrent neural networks (RNNs) for temporal-spatial feature aggregation.
• The effectiveness of RNNs in capturing temporal dependencies for VPRe-id is questioned, as they may implicitly produce orderless representations.

Purpose of the Study:

• To investigate the necessity of recurrent networks for effective visual representation learning in VPRe-id.
• To propose a novel, simpler, and powerful approach for VPRe-id that challenges the conventional RNN-based methods.

Main Methods:

• Diagnostic analysis of RNNs' performance in learning temporal dependencies for VPRe-id.
• Treating VPRe-id as an ensemble of image-based person re-identification tasks, processing videos as individual images.
• Developing an error bound under the i.i.d. assumption to guide VPRe-id improvements.

Main Results:

• Demonstrated that recurrent structures might not be as effective as expected for learning temporal dependencies in VPRe-id.
• Achieved state-of-the-art performance on multiple widely used VPRe-id datasets, including iLIDS-VID, PRID 2011, and MARS.
• Showcased a method that bridges the gap between video-based and image-based person re-identification.

Conclusions:

• Recurrent networks are not essential for learning good visual representations in VPRe-id.
• An orderless ensemble of image-based re-identification methods offers a surprisingly powerful and effective alternative.
• The proposed approach provides a promising direction for advancing VPRe-id research and applications.